case study
Rounder
Small tools, done properly: no uploads, no sign-ups, no watermarks.
01 What it is
144 tools in ten categories, from compressing a PDF to removing a background, on 212 pages counting the format and exact-size variants. Every tool runs in the visitor's browser, so files never leave their device and the site needs no server beyond one small ratings API.
02 How it works
Each engine runs in a Web Worker: WebAssembly codecs for images, pdf-lib, PDF.js and qpdf for PDFs, WebCodecs with an ffmpeg.wasm fallback for video. The 18 AI tools, among them background removal, OCR, transcription and vocal removal, run on WebGPU with a WASM fallback. Heavy jobs check the device first and warn before a phone runs out of memory.
03 The search
Type "make my pdf smaller, under 100kb" and it opens Compress PDF with the target set. A rule-based parser reads formats, sizes and platforms, keyword ranking does the rest, and a small embedding model on the device only re-ranks. A test suite gates it on held-out and adversarial queries it was never tuned on.
04 For developers and AI assistants
The same engines ship on npm as three MIT packages: an SDK, a CLI and an MCP server that gives AI assistants 29 file operations on the user's own machine. It is listed in the official MCP Registry. The packages make no network requests, and CI checks that by running their tests with outbound traffic blocked.
05 By the numbers
- 2,379 unit tests and 1,295 browser tests across Chromium, Firefox, WebKit and two phone profiles.
- CI on Linux, Windows and macOS.
- Launched in October 2026. Open source under the AGPL; the npm packages are MIT.
06 Stack
Astro, Svelte 5, Tailwind, Web Workers, WebAssembly, WebGPU, transformers.js and ONNX Runtime, Cloudflare Workers and D1.